Generative AI: A Game-Changer in Age-Related Macular Degeneration Screening by Achieving ‘Noninvasive’ Indocyanine Green Angiography

Drawbacks and Challenges of Applying Unimodal Imaging for AMD Detection

Age-related macular degeneration (AMD) is the leading cause of central vision loss in the aging population. Color fundus photography (CF) is widely applied for AMD screening due to its simplicity and low cost, particularly in source limited area. But CF images have limitations in detecting and distinguishing certain lesions, because of the unstable image quality and common characteristics shared by several chorioretinal diseases on CF images. Indocyanine green angiography (ICGA) is another well-established fundus imaging technique for screening chorioretinal conditions, owning its unique advantages in visualizing deeper choroidal vasculature and lesions behind retinal pigment epithelium. However, ICGA is an invasive imaging modality with potential adverse reactions. Besides, the complex operating procedures impede its widespread implementation in clinical settings.

Leverage the Power of GANs: Providing ‘Noninvasive’ ICGA to Enable Multimodal AMD Screening

To harness the benefits of different imaging modalities and achieve efficient multimodal AMD screening, we innovatively developed a cross-modality translation model capable of synthesizing realistic ICGA images from non-invasive and easily accessible CF images using generative adversarial networks (GANs) (Figure 1). The algorithm showed high authenticity in generating anatomical structures and pathological lesions in both internal and external datasets (Figure 2). Additionally, the integration of translated ICGA images with real CF images significantly improves the accuracy of AMD screening and effectively reduce classification errors in external validation (Table 1).

Table 1. Age-related macular degeneration (AMD) classification based on color fundus photography (CF) and CF+ translated indocyanine green angiography (ICGA) images on the AMD dataset (n=13887)

	F1-score	Sensitivity	Specificity	Accuracy	AUC	P value△
CF	0.8386	0.8368	0.9323	0.8368	0.9312
CF+early	0.8601	0.8598	0.9428	0.8598	0.9407	0.4400
CF+early+mid	0.8854	0.8850	0.9466	0.8850	0.9632	＜0.0001*
CF+early+mid+late	0.8875	0.8872	0.9474	0.8872	0.9688	＜0.0001*

Conclusion and future work

Our study pioneeringly established the feasibility of generating ICGA using CF, and introduced a cross-modality approach to augment data for AMD-related deep learning research. Furthermore, our findings highlight the potential of the CF-to-ICGA model as a valuable approach for evaluating multiple chorioretinal diseases accurately and non-invasively. Further prospective trials are required to translate this research discovery into clinical benefit in real-world practice.